See

discussions, stats, and author profiles for this publication at: http://www.researchgate.net/publication/222263261

Semidefinite programming for optimal power

flow problems
ARTICLE in INTERNATIONAL JOURNAL OF ELECTRICAL POWER & ENERGY SYSTEMS JULY 2008
Impact Factor: 3.43 DOI: 10.1016/j.ijepes.2007.12.003

CITATIONS

READS

99

151

4 AUTHORS, INCLUDING:
Xiaoqing Bai

Huang Weihua

Guangxi University

Chinese Academy of Sciences

11 PUBLICATIONS 195 CITATIONS

25 PUBLICATIONS 1,281 CITATIONS

SEE PROFILE

SEE PROFILE

Katsuki Fujisawa
Kyushu University
76 PUBLICATIONS 1,217 CITATIONS
SEE PROFILE

Available from: Xiaoqing Bai

Retrieved on: 19 November 2015

Available online at www.sciencedirect.com

Electrical Power and Energy Systems 30 (2008) 383392

www.elsevier.com/locate/ijepes

Semidenite programming for optimal power ow problems q

Xiaoqing Bai a,*, Hua Wei a, Katsuki Fujisawa b, Yong Wang c
a
College of Electric Engineering, Guangxi University, Guangxi, PR China
Department of Mathematical Sciences, Tokyo Denki University, Saitama, Japan
c
Department of Mathematics and Statistics, University of North Carolina at Charlotte, Charlotte, NC, United States
b

Received 29 April 2007; received in revised form 13 December 2007; accepted 18 December 2007

Abstract
This paper presents a new solution using the semidenite programming (SDP) technique to solve the optimal power ow problems
(OPF). The proposed method involves reformulating the OPF problems into a SDP model and developing an algorithm of interior point
method (IPM) for SDP. That is said, OPF in a nonlinear programming (NP) model, which is a nonconvex problem, has been accurately
transformed into a SDP model which is a convex problem. Based on SDP, the OPF problem can be solved by primaldual interior point
algorithms which possess superlinear convergence. The proposed method has been tested with four kinds of objective functions of OPF.
Extensive numerical simulations on test systems with sizes ranging from 4 to 300 buses have shown that this method is promising for
OPF problems due to its robustness.
2008 Published by Elsevier Ltd.
Keywords: Optimal power ow; Interior point method; Semidenite programming

1. Introduction
Since the early 60s of the last century, optimal power
ow (OPF) as a powerful tool for power system optimization problems has attracted many researchers over the
world [1,2]. Numerous algorithms have been developed in
this area based on linear programming (LP), quadratic programming, Newtons method, nonlinear programming
(NLP) and decomposition method [3]. Recently, NLPbased algorithms using interior point methods (IPM) have
also been applied to OPF problems successfully [47].
However, as power systems are getting more complex,

This work was supported in part by National Natural Science

Foundation of China under Grant 50467001, Guangxi Science Foundation under Grant 0448025, Doctoral Fund of Ministry of Education of
China under Grand 20060593002, and Guangxi Education Department
under Grant GJK No. 200420.
*
Corresponding author. Tel.: +86 771 3232264.
E-mail addresses: baixq@gxu.edu.cn (X. Bai), weihua@gxu.edu.cn (H.
Wei), fujisawa@rnc.r.dendai.ac.jp (K. Fujisawa), ywang73@uncc.edu (Y.
Wang).
0142-0615/$ - see front matter 2008 Published by Elsevier Ltd.
doi:10.1016/j.ijepes.2007.12.003

the OPF problems turn to be more dicult to handle.

Although the theory on IPM for NLP has been well developed, many issues remain open when building the links
between the modeling and the associated algorithms. First
of all, the NLP-based OPF has the convergence problem
due to its nonconvex nature. Moreover, in order to use
IPM for NLP, the Jacobian matrices (the rst-order partial
derivatives) and the Hessian matrices (the second-order
partial derivatives) have to be derived for each specic
problem. As a result, it is not convenient to develop a general and uniform software solution for the NLP problems
using IPM.
The semidenite programming (SDP) [8,9] has been one
of the most active elds in numerical optimization for over
a decade. Many well-known algorithms with uniform
frameworks have been exploited [10]. It has been proven
that the SDP is convex and the primaldual interior point
algorithms for SDP may possess superlinear convergence
theoretically [11]. Moreover, the major advantage for the
SDP-based IPM is the avoidance of deriving and computing the Jacobian matrices and the Hessian matrices for
each particular problem.

384

X. Bai et al. / Electrical Power and Energy Systems 30 (2008) 383392

Dierent practical problems have been solved successfully in various domains, such as control and system theory
[12,13], signal processing and communications [1416], and
combinatorial optimization [1719]. In [20,21], the SDP has
been used to solve power dispatch problems, which are typical constrained economic dispatching problems (CED). To
our knowledge, these two paper were pioneering in applying the SDP technique to power systems. However, the
works in [20,21] did not involve power ow equations
and the bus voltage constraints.
It is signicant and challenging to extend the SDP to
solve OPF problems which are know for their inherent
complexity and practicality. Motivations of this study stem
from the following three aspects:
1. The classical OPF problem is a nonconvex NLP [22],
which solution is more complicated than the CED problems mentioned earlier. On the other hand, the SDP
belongs to convex optimization [23], and can guarantee
global optimal solution using IPM. Therefore, it is
worth to study how to properly reformulate the classical
OPF to a SDP model, due to many advantages of the
SDP technique including its convexity and uniform
algorithm implemental framework of software suite.
2. Once an OPF problem is reformulated to a SDP one
using the quadratic model [24] in the case of the rectangular form, the resulting SDP model should be convex,
and the solution quality can be guaranteed by using
IPM for the SDP [23]. In detail, by applying X = xTx
where x is a row vector and trace operator of matrices
which will be discussed in Section 3, the nonlinear quadratic items of the OPF will be replaced with the relevant elements of the variable matrix X in SDP.
3. Mature techniques such as IPM for LP are available to
solve the SDP problems. Furthermore, the solution
techniques developed for LP which actually is a special
case of SDP are also applicable to SDP [23]. Therefore,
IPM for LP, which can be used to solve convex optimization problems in polynomial time, can be implemented for SDP [25]. Moreover, IPM enhancements
for LP can also be used in SDP, which renders IPM
for SDP as ecient as that for LP theoretically [26].

2. Formulation of OPF problem

The OPF problem is a large-scale nonlinear optimization problem. It can be formulated in polar, rectangular,
or mixture of polar and rectangular forms. In this study,
the rectangular version of the OPF problem is adopted to
take the advantages that the power ow equations are quadratic polynomials without trigonometric functions, which
can then yield the SDP models straightforward. The objective functions to be minimized are chosen out of active or
reactive power loss of transmission lines, fuel cost, and
total system active power loss. Therefore, the OPF problem
can be formulated as follows:

8
P P
2
2
P Loss2 
Gij fi  fj ei  ej 
>
>
>
i2S B j2S B
>
>
>
P P
>
>
>
Bij fi  fj 2 ei  ej 2 
< QLoss
i2S B j2S B
minimize
P
>
>
afi ali P Gi aqi P 2Gi
F Cost
>
>
>
i2S
G
>
>
P
>
>
P Gi
: P Loss1
i2S G

1
subject to:
1. Power ow equations:
8
P
>
< P Gi  j2S ei ej Gij  fj Bij fi fj Gij ej Bij  P Di
B
P
; i 2 SB
>
fi ej Gij  fj Bij  ei fj Gij ej Bij  QDi
: QRi 
j2S B

2. Constraint of reference bus:

es 1:05;

fs 0

3. Limits of active and reactive power:

P Gi 6 P Gi 6 P Gi ;

i 2 SG

QRi 6 QRi 6 QRi ;

i 2 SR

4. Limits of voltage at each bus:

V 2i 6 e2i fi2 6 V 2i ;

i 2 SB

where
a, ali, aqi: cost coecients of thermal plant, i,
respectively,
ei, fi: real and imaginary part of voltage at bus i,
V_ i : voltage at bus i,
Gij, Bij: real and imaginary part of transfer admittance
between buses i and j,
PGi, QRi: dispatchable active and reactive power at bus i,
PDi, QDi: active and reactive power demand at bus i,
SB, SG, SR: set of buses (nB), thermal plants (nG) and
reactive power sources (nB), respectively,
s: identication serial number of reference bus in the
system,
 lower and upper limits of variables or functions.
, :
The set of (1)(6) is known as the classical OPF
problem.
3. Semidenite programming by IPM
The semidenite programming [9] is concerned with
choosing a positive semidenite matrix to optimize a linear
function which is subject to linear constraints. In other
words, the well-known linear programming problem is generalized by replacing the vector of variables with a symmetric matrix and the nonnegative constraints with a positive
semidenite constraint. This generalization nevertheless
inherits several important properties from its vector

X. Bai et al. / Electrical Power and Energy Systems 30 (2008) 383392

counterpart: it is convex, has a rich duality theory

(although not as strong as the one of linear programming),
and theoretically admits ecient solution procedures based
on iterating interior points.
3.1. Semidenite programming
A semidenite programming problem can be in dierent
forms [9]. In this paper, we only consider the following
standard form [8], namely, the primal one and its dual:
8
>
< min A0  X
Primal : s:t: Ai  X bi ; i 1; . . . ; m
>
:
X 0
8
T
< max b y
m
P
Dual :
7
: s:t: A0  y i Ai  0
i1

where
1. y; b 2 Rm ; X 2 Rnn , and Ai 2 Rnn ; i 0; . . . ; m;
2. symbol P
 isP the trace operator. For example,
n
n
A1  X i1 j1P
A1ij X ji ;
m
3. X  0P and A0  i1 y i Ai  0 mean that X and
m
A0  i1 y i Ai are positive semidenite, respectively.
3.2. IPM for SDP
All the improvements developed for IPM as applied to
linear programming are also useful when they are extended
to SDP. For example, the primaldual interior point
method (PDIPM) is used to solve the SDP problem successfully [8,27]. More clearly, PDIPM generates a sequence
of iterations which approximate the central path and converge to the primal and dual solutions [28].
3.2.1. Primaldual interior point algorithms
Let ln det X be the logarithmic barrier for the determinant of the matrix X. Replacing the primal problem in
(7) with the relaxed barrier version [8] yields:

min A0  X  l ln det X
8
s:t: Ai  X bi ; i 1; . . . ; m
The Lagrangian function is then given by:
m
X
y i bi  Ai  X
LX ; y  A0  X  l ln det X 
i1

Applying the rst-order KarushKuhnTucker (KKT)

optimality conditions leads to:
8
m
< rLl A0  P Ai y l  lX l1 0
X
i
i1
:
l
rLyi bi  Ai  X l 0; i 1; . . . ; m
Now introducing a symmetric matrix Z yields the following
nonlinear system:

385

8
Ai  X l bi ; i 1; . . . ; m
>
>
<P
m
y i lAi Zl A0
>
i1
>
:
X lZl lI

where the last equation is equivalent to Z(l) = lX(l)1.

Then, the same approach above is applied to the dual:
8
T
< max b y  l ln det Z
m
P
10
: s:t:
y i Ai Z A0
i1

The dual Lagrangian now is (with the Lagrange multiplier

being a symmetric matrix):
!
m
X
T
LX ; y; Z b y  l ln det Z  X  A0 
y i Ai  Z
i1

Applying the KKT optimality conditions, we get:

8
m
P
l
>
>
>
< rLX A0  Ai y i l  Zl 0
i1

rLlyi bi  Ai  X l 0; i 1; . . . ; m
>
>
>
:
rLlZ mZ 1 l  X l 0
which is essentially the same as the one obtained from the
primal problem.
3.2.2. Central path
In PDIPM [29], the central path (l) is composed of
points dened by system (9):
 l  fX l; yl; Zl 2 Rnn  Rm  Rnn : l > 0g
The central path plays a crucial role in computing an optimal solution. For any complementary gap l > 0, there exists
a
unique
X(l), y(l), Z(l) 2 Rnn  Rm  Rnn
satisfying (10). Thus the central path (l) structures a
smooth curve in the space of X(l), y(l), Z(l) 2
Rnn  Rm  Rnn. Moreover, it is obvious that
X(l)Z(l) = nl can be held since X(l)Z(l) = lI. Therefore, l = X(l)Z(l)/n. In addition, the central path (l)
converges to an optimal solution of the SDP, i.e.,
(X(l), y(l), Z(l)) converges to an optimal solution of the
SDP as l ? 0. Thus, the PDIPM traces path numerically
as l decreases toward 0.
3.2.3. Search direction
A search direction (DX, Dy, DZ) [30] is taken so that
(XK + DX, yK + Dy, ZK + DZ) can coincide with the target
point (X(blk), y(blk), Z(b lk)) on the central path with
l = blk, which leads to the following system by applying
Newtons method to (9):
8
Ai  X k DX bi ; i 1; . . . ; m
>
>
>
<P
m
y ki DyAi Z k DZ A0
11
>
i1
>
>
:
X k DX Z k DZ blk I

386

X. Bai et al. / Electrical Power and Energy Systems 30 (2008) 383392

System (11) is almost a linear system except the term DXDZ

in the last equality. Neglecting the nonlinear term and
e , the system of nonlinintroducing an auxiliary matrix D Z
ear equations above can be reduced into the following system of linear equations:

12

i1

>
>
>
DZ~ X k 1 R  Z k DX
>
>
:
e DZ
e T =2
DZ D Z

i; j 1; . . . ; m

k 1

ri d i Ai  X R  PZ k ;
m
X
P
Ai y ki  A0  X k

i 1; . . . ; m

i1

d i c i  Ai  Z k ;

3.2.4. The framework of the PDIPM algorithm for SDP

The general framework of the PDIPM algorithm for
SDP is described as follows:
Initialization: set the tolerance e = 105, the centering
parameter b = 0.2 and the step length reduction factor
c = 0.85. Note that they actually can be held by 0 < b < 1
and 0 < c < 1. The more detailed discussion about how to
properly choose b and c is presented in Part E of Section
4. Choose the initial point (X0, y0, Z0) with X0  0, Z0  0.
Begin k = 0, 1,. . .;
Step 1: lk = XkZk/n. If lk < e and (Xk, yk, Zk) approximately satises feasibility, (Xk, yk, Zk) as an approximate
optimal solution is printed out and stop;
Step 2: A search direction (DX,Dy, DZ) is computed
toward a target point (X(blk), y(blk), Z(blk)) on the
central path (l) with l = blk by solving (12).
Step 3: The step lengths ap and ad are computed such
that Xk + apDX and Zk + adDZ to Xk+1 and Zk+1
remain in the interior of the cone of positive denite
matrices. Let L be the lower triangular matrix from
the Cholesky factorization of Xk = LLT and let PKPT
be the eigenvalue decomposition of L1D XLT.
Then
0<c<1

4. Implementation of SOPF

4.1. Process of SOPF

i 1; 2; . . . ; m

R blk I  X k Z k

ap g
ap c minf1; 

14

The solution using the SDP technique to solve the

OPF problem is called SOPF. In principle, the OPF
problem can be reformulated into SDP shown in (7)
using its quadratic style, and then solved by PDIPM.
Afterward, the solution of the classical OPF can be
mapped from the solution of SDP for OPF in a
straightforward way. More detailed discussions about
SOPF are presented below.

where
1

X k1 ; y k1 ; Z k1 X k ap DX ; y k ad Dy; Z k ad DZ
End

8
BDy r
>
>
>
m
>
P
>
< DX P Ai Dy i

Bij X k Ai Z k   Aj ;

Step 4: Update the current point by

13

where:

1=kmin ; if kmin < 0

ap
1;
otherwise
Here kmin is the minimum diagonal value of K. The set
lengths ad can be computed in the same way.

Once the relationship between the classical OPF

problem and OPF in SDP style is conrmed by the
analysis in the later two subsections, the left task is
as follows:
Step 1: Set data to appropriate places in A which are
coecient matrices in (31) according to the analysis in
Section 4.1.1 and 4.1.2.
Step 2: Invoke IPM solver for SDP as presented in Section 3.
Step 3: Map solutions to the classical OPF from
the result of Step 2 to be described in Section
4.1.3.

4.1.1. OPF formulated as quadratic problem

For the classical OPF problem shown in (1)(6), the
object functions (FCost, PLoss2, and QLoss) are quadratic,
and (PLoss1) is linear, respectively; all the constraints are
quadratic or linear.
Let dGi(i 2 SG) and dRi (i 2 SR) be auxiliary variables
which are always equal to one, i.e., dGi = 1 and dRi = 1.
Obviously, there exists d 2Gi 1 and d 2Ri 1 since dGi = 1
and dRi = 1. On the other hand, also exists PGi = PGidGi
and QRi = QRidRi. Therefore, each linear term about
PGi and QRi in the classical OPF problem can be reformulated into the quadratic terms about PGi nd QRi with PGidGi
and QRidRi, respectively.
Meanwhile, by introducing a slack variable vector
uG ; lG ; uR ; lR ; uB ; lB  2 R2nG 2nR 2nB , inequality constraints
(4)(6) can be transformed into the equality constraints
as u2G ; l2G ; u2R ; l2R ; u2B ; l2B P 0. Thus, the classical OPF
(1)(6) can be equivalently formulated as a quadraticobjective and quadratic-constraint problem (15)(28) as
follows:

X. Bai et al. / Electrical Power and Energy Systems 30 (2008) 383392

"
#
8
nB
nB
nB
P
P
P
>
2
2
>
>
P


2f

G

G

4
f
f

e
e
G
Loss2
ij
ii
i
j
i
j
ij
>
i
i
>
>
i1
j1
ji1
>
>
"
#
>
>
>
nB
nB
nB
>
P
P
P
<
2
2
2fi ei Bij  Bii  4
fi fj ei ej Bij
QLoss
minimize
i1
j1
ji1
>
>
P
>
>
>
F Cost
afi d 2Gi ali P Gi d Gi aqi P 2Gi
>
>
>
i2S G
>
>
P
>
>
: P Loss1
P Gi d Gi

G2

387

G1

(Reference)

Fig. 1. TEST-4: a simple power system.

i2S G

15

sources, the buses 2 and 3 have scheduled active and reactive power, and the bus 4 is the reference bus.
All the variables in the problem (15)(28) for TEST-4
are grouped into a single vector x = [x1, x2, x3, x4, x5, x6],
where

subject to:
1. Power ow equations:
XnB
ei ej Gij ei fj Bij  fi fj Gij  fi ej Bij
j1
P Di ; i 2 S B =S G
XnB
fi ej Gij fi fj Bij ei fj Gij ei ej Bij
j1

16

QDi ; i 2 S B =S R
nB
X
P Gi d Gi
ei ej Gij ei fj Bij  fi fj Gij  fi ej Bij

17

x1 P G1

18

i 2 S R

20

1:05

21

P Gi d Gi u2Gi P Gi ;

i 2 SG

22

P Gi d Gi  l2Gi
QRi d Ri u2Ri
QRi d Ri  l2Ri

P Gi ;

i 2 SG

23

QRi ;

i 2 SR

24

QRi ;

i 2 SR

25

4. Limits of amplitude at each bus:

e2i

fi2

l2Bi

2
i;

i 2 SB

26

i 2 SB

27

5. Auxiliary variables
d 2Gi 1;

i 2 SG

d 2Ri

i 2 SR

1;

x4 uR1

x5 e1

uG2

lG2 

d R1

QR2

d R2 

lR1

uR2

lR2 

28

4.1.2. The OPF in quadratic form reformulated as SDP

Now the classical OPF problem has been formulated as
a quadratic problem (15)(28) which can be transformed
into a SDP problem of form (7). For simplicity, TEST-4,
a simple power system shown in Fig. 1 is used to explain
the data matrix structures, where the generators G1 and
G2 at the buses 1 and 4, respectively, are dispatchable

f1

e2

f2

e3

f3

e4

f4 

6. Group of slack variables for buses:

x6 uB1

3. Limits of active and reactive power:

e2i fi2 u2Bi V 2i ;

lG1

5. Group of buses:

2. Constraint of reference bus:

e2s

d G2 

4. Group of slack variables for reactive resources:

19

fs2 0

P G2

3. Group of reactive resources:

x3 QR1

j1

QDi ;

d G1

2. Group of slack variables for active resources:

x2 uG1

j1

P Di ; i 2 S G
nB
X
fi ej Gij fi fj Bij ei fj Gij ei ej Bij
QRi d Ri

1. Group of active resources:

lB1

uB2

lB2

uB3

lB3

uB4

lB4 :

Afterward, the SDP variable can be dened by:

X xT x
29
nn
Obviously, the matrix X 2 R is positive denite or semidenite [31], i.e., X  0. And in this example, it has dimension n = 2  2  nG + 2  2  nR + 2  2  nB = 32.
The matrix X is shown in (30). For simplicity, some elements in X are replaced with ellipses, indicating that the
relevant coecients of those elements are always zeroes.
Therefore, the matrix X can be treated as a block-diagonal
symmetric matrix. This novel block-diagonal symmetric
matrix variable X can take advantage of the sparsity technique of SDP [32], so that the computing eciency can be
improved to some extent.
3
2
X1


7
6 .
..
7
6 .. X 2 . . .
.
7
6
7
6
.
.
.
6
.. X
.. 7
..
7
6
3
7
X 6
30
..
..
7
6 ..
7
. X4
6 .
.
7
6
..
..
.. 7
6
7
6
.
X
.
.
5
5
4
..


. X 6 3232

388

X. Bai et al. / Electrical Power and Energy Systems 30 (2008) 383392

where
2

P 2G1

6
6 P G1 d G1
X1 6
6P P
4 G1 G2
P G1 d G2
2

u2G1

P G1 d G1

P G1 P G2

d 2G1

P G2 d G1

P G2 d G1

P 2G2

d G1 d G2

P G2 d G2

uG1 lG1

uG1 uG2

6u l
6 G1 G1
X2 6
4 uG1 uG2

l2G1

lG1 uG2

lG1 uG2

u2G2

uG1 lG2

lG2 lG1

lG2 uG2

Q2R1

QR1 d R1

P G1 d G2

7
d G1 d G2 7
7
P G2 d G2 7
5
d 2G2
44
3
uG1 lG2
lG2 lG1 7
7
7
lG2 uG2 5
l2G2

QR1 QR2

The structures as well as more detailed discussions about

the construction of the matrices Ai (i = 0, 1, . . ., 30) are described as follows:
3
2
0

0
W Gi
..
7
6 0
W SGi
0
.
6
.. 7
7
6
6
0
. 7
0
W Ri
Ai 6 .
7
32
7
6 ..
0
W SRi
0
7
6
..
5
4
.
0
W Bi
0
0

0
W SBi 3232

44

QR1 d R2

7
6
6 QR1 d R1
d 2R1
QR2 d R1 d R1 d R2 7
7
X3 6
6Q P
2
QR2
QR2 d R2 7
5
4 R1 R2 QR2 d R1
QR1 d R2 d R1 d R2 QR2 d R2
d 2R2
44
3
2 2
uR1
uR1 lR1 uR1 uR2 uR1 lR2
6u l
l2R1
lR1 uR2 lR2 lR1 7
7
6 R1 R1
X4 6
7
4 uR1 uR2 lR1 uR2
u2R2
lR2 uR2 5
2

uR1 lR2
e21

6
6 e1 f1
6
6 e1 e2
6
6
6 e1 f2
X5 6
6e e
6 1 3
6
6 e1 f3
6
6
4 e1 e4
2

e1 f4

lR2 lR1

lR2 uR2

l2R2

44

e1 f1

e2 e1

f2 e1

e3 e1

f3 e 1

e4 e1

f12

e 2 f1

f2 f1

e3 f1

f3 f1

e4 f1

f1 e2

e22

f2 e2

e3 e2

f3 e 2

e4 e2

f1 f2

e 2 f2

f22

e3 f2

f3 f2

e4 f2

f1 e3

e2 e3

f2 e3

e23

f3 e 3

e4 e3
e4 f3

f1 f3

e 2 f3

f2 f3

e3 f3

f32

f1 e4

e2 e4

f2 e4

e3 e4

f3 e 4

e24

f1 f4

e 2 f4

f2 f4

e3 f4

f3 f4

e4 f4

u2B1
6
6 uB1 lB1
6
6 uB1 uB2
6
6
6 uB1 lB2
X6 6
6u u
6 B1 B3
6
6 uB1 lB3
6
6
4 uB1 uB4

lB1 uB1

uB2 uB1

uB1 lB4

lB2 uB1

uB3 uB1

f4 e1

1. Construction of A0
For simplicity, the object function of minimizing the fuel
cost (FCost) is shown here, and the other three object
functions (PLoss2, QLoss, and PLoss1) can be constructed
in the same way. With the denition for X in (30) and
the structure of Ai in (32), A0 is
2
3
W G0 0 0
6
.7
..
6 0
0
. .. 7
7
A0 6
.7
.. ..
6 ..
4 .
. .. 5
.
0
0 3232

7
f4 f1 7
7
f4 e2 7
7
7
f4 f2 7
7
f4 e3 7
7
7
f4 f3 7
7
7
f4 e4 5
f42 88
lB3 uB1

uB4 uB1

l2B1

uB2 lB1

lB2 lB1

uB3 lB1

lB3 lB1

uB4 lB1

lB1 uB2

u2B2

lB2 uB2

uB3 uB2

lB3 uB2

uB4 uB2

lB1 lB2

uB2 lB2

l2B2

uB3 lB2

lB3 lB2

uB4 lB2

lB1 uB3

uB2 uB3

lB2 uB3

u2B3

lB3 uB3

uB4 uB3

lB1 lB3

uB2 lB3

lB2 lB3

uB3 lB3

l2B3

uB4 lB3

lB1 uB4

uB2 uB4

lB2 uB4

uB3 uB4

lB3 uB4

u2B4

lB1 lB4

uB2 lB4

lB2 lB4

uB3 lB4

lB3 lB4

uB4 lB4

where
lB4 uB1

7
lB4 lB1 7
7
lB4 uB2 7
7
7
lB4 lB2 7
7
lB4 uB3 7
7
7
lB4 lB3 7
7
7
lB4 uB4 5
2
lB4
88

With the denition and the structure of X, the problem

(15)(28) for TEST-4 can be easily reformulated as the primal SDP of (7) as following:
min

F A0  X

s:t:

Ai  X b i ;
X 0

i 1; . . . ; 30

31

where the coecient matrices Ai have the same dimension

as X, and the relationships between i and the constraints
are below:
A1 ; A2 () Constraint 16; A3 ; A4 () Constraint 17;
A5 ; A6 () Constraint 18; A7 ; A8 () Constraint 19;
A9 () Constraint 20; A10 () Constraint 21;
A11 ; A12 () Constraint 22; A13 ; A14 () Constraint 23;
A15 ; A16 () Constraint 24; A17 ; A18 () Constraint 25;
A19 ; A20 ; A21 ; A22 () Constraint 26;
A23 ; A24 ; A25 ; A26 () Constraint 27;
A27 ; A28 ; A29 ; A30 () Constraint 28:

W G0

aq1
61a
6 l1
62
4 0

1
a
2 l1

0
0

af 1
0

aq2

1
a
2 l2

3
0
0 7
7
7
1
5
a
l2
2
af 2

44

2. Construction of Ai, i = 1, . . ., 30
The constraints matrices are constructed in a similar
way to A0s. For illustration purpose, we describe the
derivations of A5 for (18) which originally is an equation
constraint, A11 for (22) which originally is an inequality
constraint, and A27 for (28) which is the auxiliary variable constraints, respectively.
(1) A5
The constraint (18) is the active power ow equation of
the ith generator in the system. For the generator G1 in
TEST-4, it is A5, and there exists:
A5  X P G1 d

4
X
e1 ej G1j e1 fj B1j  f1 fj G1j  f1 ej B1j
j1

P D1

More precisely, the structure of A5 is

3
2

0
W G5 0
..
7
6
7
6 0
0
.
6
.. 7
..
7
6
6
0
. .7
A5 6 .
7
..
7
6 ..
.
0
0
7
6
7
6
..
4
. 0 W B5 0 5
0

0
0 3232

X. Bai et al. / Electrical Power and Energy Systems 30 (2008) 383392

where
2

16
61 0
W G5 6
240 0

0
0

07
7
7
05

W B5

Other constraints in the OPF quadratic form can be

reformulated into AiX = bi in the same way as that
for A5, A11, or A27.
3. Construction of bi, i = 1, . . ., 30
The vector b is constructed using the constants located
in the right hand side of the constraints (16)(28) in
TEST-4, namely,

0 1

0 0 0 0 44
2
0
G12
2G11
6 0
2G
B
11
12
6
6
6 G12 B12
0
6
6
6 B
G12
0
12
1 6
6
 6
2 6 G13 B13
0
6
6
6 B
G13
0
6 13
6
6
4 G14 B14
0
G14
0
B14

B12
G12

G13
B13

B13
G13

G14
B14

0
..
.
..
.
..
.
..
.
0

..
.

..

.
..

3
B14
G14 7
7
7
0 7
7
.. 7
. 7
7
.. 7
7
. 7
7
.. 7
. 7
7
.. 7
7
. 5
0

bT P D2 P D3 QD2 QD3 P D1 P D4 QD1 QD4 0 1:052

P G1 P G2 P G1 P G2 QR1 QR2 QR1 QR2 . . . . . .
V 21 V 22 V 23 V 24 V 21 V 22 V 23 V 24 1 1 1 1
4.1.3. Mapping solutions to the classical OPFs
88

(2) A11
The constraint (22) is the maximum active power of the
ith generator in the system. For the generator G1 in
TEST-4, there exists:
A11  X P G1 d u2G1 P G1
More precisely, the structure of
2
0

W G11
6
..
6 0
.
W SG11
6
6 .
A11 6
0
6 ..
6
..
..
6 ..
4 .
.
.

0
where

W G11

3
0 1 0 0
7
16
61 0 0 07
6
7
240 0 0 05
0 0 0 0 44

A11 is
3
0
7
7
7
.. 7
7
.7
7
7
5
0

W SG11

1 0 0
60 0 0
6
6
40 0 0
0 0 0

3
0
07
7
7
05
0

The solution of OPF in SDP (31) can be mapped to the

solution of the classical OPF (1)(6) in a straightforward
way. more precisely, upon obtaining an optimal solution
X* of (31) and referring to the structure of X in (30), the
optimal solution of (1)(6) can be obtained in many ways
since X* contains enough information. The following procedure is for the purpose of illustration:
1. Active power PGi(i 2 SG) and reactive power QGi
(i 2 SR) equal to the square roots of the corresponding
diagonal elements P 2Gi and Q2Ri in the blocks X1 and X3
of (30), respectively.
2. Real part ei(i 2 SB) of voltage equal to the square roots
of the corresponding diagonal elements e2i in the block
X5 of (30). As for imaginary part fi(i 2 SB) of voltage,
fi = eifi/ei, where eifi is the right-hand adjoining element
to e2i in X5 of (30).
In this mapping procedure, the elements of the matrix
X* mentioned above are more than enough to obtain the
solution of (1)(6). Therefore, no all elements are involved.

3232

389

44

(3) A27
The constraint (28) is for the auxiliary variables which
are equal to one. For the generator G1 in TEST-4, it
is A27, and since dG1 = 1, there exists d 2G1 1. Thus
the structure of A27 is

4.2. Choice of the initial value

In general, the better the initial values are selected, the

faster the convergence of an algorithm will be. In [33],
the initial values of the OPF variables are set as the averages of their limits which are feasible. However, it has been
observed that the proposed algorithm in this study is not
quite sensitive to the initial values whether they are feasible
or unfeasible [28]. Experimentally, it works very well that
the initial values are given by data whose order of magnitude is the same as that of an optimal solution
(X*, y*, Z*). Both ways for choosing the initial values lead
to almost the same computing eciency in this study.
Therefore, in this study, the initial values do not have to
be set scrupulously. Instead, for simplicity and without loss
of any generality, they are chosen as follows:

60
6
W G27 6
40

1
0

0
0

07
7
7
05

1. The primal variable matrix X(0) and the slack variable

matrix Y(0)

3
0
.7
..
0
. .. 7
7
.. 7
.. ..
7
. .5
.
0 3232
0

W G27

6
6 0
6
A27 6 .
6 .
4 .
0
where
2

44

X 0 s  I;

Y 0 s  I

390

X. Bai et al. / Electrical Power and Energy Systems 30 (2008) 383392

where is a coecient whose order of magnitude is the

same as that of the optimal solution (X*, y*, Z*), and I
denotes the identity matrix which has the same dimension as the matrix X or Y in (30).
2. The dual variable vector Y(0)
y 0 0
where 0 denotes a full zero vector which has the same
dimension as the vector formed by bi in (31).

Table 2
Solutions by SDP and NLP (PER UNIT)
By SDP

Active
Power
Reactive
Power
Voltage

5. Test results and discussions

Node

Value

Node

Value

1
4
1
4
1
2
3
4

0.56818
0.30000
0.10048
0.27613
0.98513 + 0.01257i
0.95976  0.09846i
1.08660 + 0.17117i
1.05 + 0i
421.7741

1
4
1
4
1
2
3
4

0.56818
0.30000
0.10048
0.27613
0.98513 + 0.01257i
0.95976  0.09846i
1.08660 + 0.17117i
1.05 + 0i
421.7741

Fuel cost($)

The proposed algorithm has been implemented in

SDPAM [30] using Matlab 7.0 on a 2.8 GHz Pentium 4,
Windows PC with 512 MB of RAM for six dierent test
systems. The details of the six test systems, including a simple 4-bus system and ve IEEE standard systems, are
shown in Table 1. It should be noted that the OPF problem
in SDP formulation can be solved with only two lines of
Matlab codes using SDPAM, which actually benets from
the general uniform framework of the SDP software suite.
5.1. Solution quality of OPF using SDP
The OPF problem can be solved by IPM for nonlinear
programming with excellent computing performance, and
its solution is optimal [24]. However, the Jacobian matrices
and the Hessian matrices have to be derived for each specic problem using NLP. Therefore, it is not convenient
to develop a general uniform software solution for OPF
problems using NLP.
On the other hand, the OPF problem can also be solved
by IPM for semidenite programming, and the numerical
simulation in this study has demonstrated that its solution
quality can be guaranteed like as for NLP. For the sake of
space, we list in Table 2 only the results of TEST-4 by SDP
and NLP which, as can be seen, are identical.
Table 3 shows the CPU time and the number of iterations for the optimal solutions by SOPF for the four kinds
of objective functions of all six test systems.
It should be point out that OPF using SDP can not compete against OPF using NLP in terms of CPU times so far.
Therefore, it is worthwhile to explore more ecient algorithms for SDP in the future.

Table 3
CPU time (s), iterations
Name of systems

PLoss2

QLoss2

FCost

PLoss1

TEST-4
IEEE-14
IEEE-30
IEEE-57
IEEE-118
IEEE-300

0.0156/9
0.0781/15
0.3906/17
2.2813/24
12.72/24
121.38/31

0.0156/11
0.0781/14
0.3438/15
1.7656/18
12.781/25
122.90/27

0.0156/9
0.0781/9
0.3125/12
1.6406/19
9.0938/22
102.05/26

0.0156/10
0.0938/14
0.3594/25
1.2188/17
6.9219/21
154.31/21

5.2. Convergence of OPF by SDP

The complementary gap is a very important measure to
judge the optimality of solutions, and its change reects the
characteristic of the algorithm. Fig. 2 shows how it reduces
to zero (here, tolerance is e = 105) with iterations for the
case IEEE-300 under four dierent kinds of objective functions. Generally speaking, the complementary gap should
decrease to zero with iterations monotonously and rapidly
for an algorithm to be ecient. For this reason, Fig. 2
clearly demonstrates the eciency of the proposed algorithm. As for the remaining important feature of the algorithm, Fig. 3 shows the relative optimal values of four
objective functions converge with iterations.
Just like IPM for NLP, tuning the centering parameter b
introduced in (12) and the reduction factor c introduced in
(13) of IPM for SDP can exploit the tradeo between the
accuracy of optimal solutions and the computation times.

10

Name of
system
TEST-4
IEEE-14
IEEE-30
IEEE-57
IEEE-118
IEEE-300

Number of
buses/lines
4/4
14/20
30/45
57/78
118/179
300/409

Controllable/
state variables
(PGi,QRi/ei,fi)

Slack
variables

1, 2/4, 4
5, 3/14, 14
6, 6/30, 30
4, 7/57, 57
16, 54/118, 118
21, 69/300, 300

14
44
84
136
376
780

Number of
constrains
27
75
147
253
615
1383

Complementary Gap

10

Table 1
Test systems size, control variables and inequality constraints

By NLP

10
10
10
10
10

min Ploss2
min Qloss
min Fcost
min Ploss1

-2

-4

-6

10

15
20
Iterations

25

30

35

Fig. 2. Complementary gaps with iterations for IEEE-300 system.

X. Bai et al. / Electrical Power and Energy Systems 30 (2008) 383392

Object Function Values

10

10

10

10

10

References

min Ploss2
min Qloss
min Fcost
min Plost1

391

10

15
20
Iterations

25

30

35

Fig. 3. Objective functions with iterations for IEEE-300 system.

The related discussion, however, is not included in this

paper again due to space limitation.
6. Conclusions and future work
The solution to the OPF problem using semidenite programming has been presented in this paper. The proposed
method has been tested on six test systems with sizes ranging from small (4 buses) to large (300 buses), and the results
are satisfactory. To the best of our knowledge, the SDP
technique has not been used previously to solve the OPF
problem.
The OPF is a very complex nonlinear and nonconvex
problem which has been solved by various optimization
methods. IPM is one of the best algorithms for the OPF
problem in NLP forms [8]. The SDP is a technique for convex programming that can be implemented using IPM as
well. Therefore, once a classical OPF problem is transformed into the semidenite programming model, it turns
to be a convex problem, and then can be solved using
IPM for SDP and benets from the SDP technique, as
shown in the extensive numerical simulations presented in
this paper.
However, the proposed method could not solve extremely large systems due to limitation in computer resources
such as CPU and memory. Therefore, developing more
powerful algorithms with sparsity technique and parallel
computing technique for OPF using SDP will be an attractive research topic in the future.
Acknowledgements
The authors wish to thank Professor Masakazu Kojima
from Department of Mathematical and Computing Sciences, Tokyo Institute of Technology and his SDPA project team. Thanks also go to Dr. Makoto Yamashita,
member of the SDPA team from Department of Industrial
Engineering and Management in Kanagawa University,
Japan, for discussion on semidenite techniques and on
SDPAM which is one of the best MATLAB-based SDP
solvers in the world [10].

[1] Carpentier J. Contribution a letude de Dispatching Economique. Bull

Soc Fr Electric 1962;3:43147.
[2] Dommelll HW, Tinney WF. Optimal power ow solutions. IEEE
Trans PAS 1968;87:186676.
[3] Momoh JA, Koessler RJ, Bond MS, Stott B, Sun D, Papalexopoulos
A, et al. Challenges to optimal power ow. IEEE Trans Power Syst
1997;12:44455.
[4] Granville S. Optimal reactive dispatch through interior point
methods. IEEE Trans Power Syst 1994;9:13646.
[5] Wei H, Sasaki H, Yokoyama R. An application of interior point
quadratic programming algorithm to power system optimization
problems. IEEE Trans Power Syst 1996;11:2607.
[6] Wu Y-C, Debs AS, Marsten RE. A direct nonlinear predictorcorrector primaldual interior point algorithm for optimal power
ows. Presented at power industry computer application conference,
Scottsdale, AZ, USA; 1993.
[7] Quintana VH, Torres GL, Medina-palomo J. Interior-point methods
and their applications to power systems: a classication of publications. IEEE Trans Power Syst 2000;15:1706.
[8] Adler I, Alizadeh F. Primaldual interior point algorithms for convex
quadratically constrained and semidenite optimization problems.
Technical report, Rutcor, Rutgers University, New Brunswick, NJ;
1995.
[9] Todd MJ. Semidenite optimization. Acta Numer 2001;10:
51560.
[10] Mittelmann HD. An independent benchmarking of SDP and SOCP
solvers. Math Program 2003;B:40730.
[11] Wolkowicz H. Handbook of applied optimization. Oxford University
Press; 2001.
[12] P. Parrilo. Structured semidenite programs and semialgebraic
geometry methods in robustness and optimization. PhD thesis.
California: Caltech, Pasadena; 2000.
[13] Parrilo P. Semidenite programming relaxations and algebraic
optimization in control. Eur J Control 2003;9:30721.
[14] Alkire B, Vandenberghe L. Handling nonnegative constraints in
spectral estimation. Technical report, Electrical Engineering Department, UCLA, Los Angeles, CA; 2000.
[15] Alkire B, Vandenberghe L.Convex optimization problems involving
nite autocorrelation sequences. Technical report, Electrical Engineering Department, UCLA, Los Angeles, CA; 2001.
[16] Xiao L, Boyd S. Fast linear iterations for distributed averaging.
Technical report, Stanford University, Stanford, CA; 2003.
[17] Benson SJ, Ye Y, Zhang X. Solving large scale sparse semidenite
programs for combinatorial optimization. SIAM J Optim
2000;10:44361.
[18] Goemans MX, Rendl F. Semidenite programming in combinatorial
optimization. Math Program 1997;79:14361.
[19] Goemans MX. Semidenite programming and combinatorial optimization. Documenta mathematica, Extra Volume ICM 1998. p.
65766.
[20] Fuentes-Loyola R, Quintana VH. Medium-term hydrothermal coordination by semidenite programming. IEEE Trans Power Syst
2003;18:151522.
[21] Madrigal M, Quintana VH. Semidenite programming relaxations
for {0, l}-power dispatch problems. Presented at proceeding of IEEEPES 1999 Summer Meeting, Edmonton, Al., Canada; 1999.
[22] Jabr RA. A primaldual interior-point method to solve the
optimal power ow dispatching problem. Optim Eng 2003;4:30936.
[23] Alizadeh F, Haeberly J-PA, Overton ML. Primaldual interior-point
methods for semidenite programming: convergence rates, stability
and numerical results. Computer Science Department, Technical
Report 721, New York University, May 31; 1996.
[24] Wei H, Sasaki H, Kubokawa J, Yokoyama R. A interior point
nonlinear programming for optimal power ow problems with a
novel data structure. IEEE Trans Power Syst 1998;13:8707.

392

X. Bai et al. / Electrical Power and Energy Systems 30 (2008) 383392

[25] Mizuno S, Kojima M, Todd MJ. Infeasible-interior-point primal

dual potential reduction algorithms for linear programming. SIAM J
Optim 1995;5:5267.
[26] Alizadeh F. Combinatorial optimization with interior point methods
and semidenite matrices. PhD Thesis, University of Minnesota; 1991.
[27] Wright S. Primaldual interior-point methods. Publisher: Soc for
Industrial & Applied Math, June 1997.
[28] Fujisawa K, Fukuda M, Kojima M, Nakata K. Numerical evaluation
of SDPA (Semidenite Programming Algorithm). Technical report
series B: operations research, Department of Mathematical and
Computing Sciences, Tokyo Institute of Technology, Tokyo, September 1998.
[29] Yamashita M, Fujisawa K, Kojima M. SDPARA: SemiDenite
programming algorithm PARAllel version. Parallel Comput
2003;29:105367.

[30] Satoshi Matsuyama SN, Katsuki Fujisawa, Kazuhide Nakata,

Masakazu Kojima. SDPA-M (SemiDenite Programming Algorithm
in MATLAB) users manual Version 1.00. Technical report series
B: operations research, Department of Mathematical and
Computing Sciences, Tokyo Institute of Technology, Tokyo, January
2000.
[31] Lovasz L, Schrijver A. Cones of matrices and set-functions and 01
optimization. SIAM J Optim 1991;1:16690.
[32] Fujisawa K, Kojima M, Nakata K. Exploiting sparsity in primal
dual interior-point methods for semidenite programming. Math
Program 1997;79:23553.
[33] Wei H, Sasaki H, Kubokawa J, Yokoyama R. Large scale
hydrothermal optimal power ow problems based on interior
point nonlinear programming. IEEE Trans Power Syst 2000;15:
396403.